Graph-Grounded Retrieval-Augmented Generation for S2B2C Industrial Safety: A Knowledge-Centric and Auditable Framework

Industrial safety management increasingly relies on decision support that must operate across organizations with different procedures, incentives, and regulatory exposure.We present a Graphgrounded Retrieval-Augmented Generation (G-RAG) method for S2B2C safety platforms, where the platform layer maintains shared knowledge and governance, enterprise units tailor it to assets and jurisdictions, and frontline crews execute task-level procedures.The method combines a Safety Knowledge Graph (Safety-KG) with text corpora and operational signals.It retrieves graph paths and text spans jointly, checks SHACL-like constraints before generation, and selects a budget-feasible control set that targets expected risk reduction while maintaining coverage across control layers.In a simulation study on large synthetic benchmarks, G-RAG improves evidence retrieval (nDCG@10), increases evidence sufficiency for generated claims, reduces compliance violations, and achieves higher expected risk reduction under fixed budgets than text-only and rule-based baselines.The output is audit-ready by design, with each control-level claim linked to a compact set of supporting paths and passages with provenance.